The $300 million Gates refused, and the $10 billion that went to OpenAI

Nathan Labenz opens by naming a role. He calls it the AI scout, after Julia Galef's *The Scout Mindset*: the work of forming accurate beliefs even when they are uncomfortable. Models still fall short of human experts, he says, but they already beat the average person on a long list of economically useful tasks, and they keep improving. Their architecture is unlike ours, their strengths and failure modes are unlike ours, and their behavior is hard to predict. Society, in his view, is flying fast and half-blind. Governments and firms already fund intelligence on rivals and markets. It is only prudent, he argues, to fund people who study AI from every angle, at every level of abstraction, so the rest of us can decide how to build and use it on the basis of what it actually is.

He tries to spend about half his time on that work: papers, products, models, Twitter, other podcasts, and the builders who come on this show. The aim is no major blind spots, plus enough depth to help teammates, advisees, the audience, and maybe policymakers. Riley Goodside, another early scout, said on a previous episode that everyone is still new. The pace is relentless and the firehose is real, but that same speed is why a newcomer can still reach the frontier and start contributing. This episode is two other scouts, working in their own ways: Robert Scoble and Ben Tossell.

Scoble's résumé sounds like a Silicon Valley shaggy-dog story until you notice it is a method. At thirteen he was, in his telling, Apple's first child laborer, the benefit of an electrical-engineer father who moved the family to Cupertino in 1971. He and his mother stuffed a couple hundred Apple II motherboards. Apple was his first startup. He was early on Siri and Flipboard, the 79th user of Instagram, first passenger in Elon Musk's first Tesla, first rider in Mercedes-Benz's first autonomous car. He has written four books that try to predict decade-scale change; the last is on spatial computing, which is why he is watching AI. On Twitter he follows 38,000 people in the field and claims to be the only human doing that, because he noticed Stanford and Carnegie Mellon students grouping up to ship new models and papers at a pace he could not cover any other way.

He still uses Twitter less as a broadcast than as a sensor: the market, his investments, the papers. Microsoft, he says, now has data centers a mile long. He used to work at Rackspace and has been inside a lot of those buildings; the new ones are a different species, as are the data flows. ChatGPT is not Siri. It writes code, then explains the code, then writes marketing copy and email. It is still next-word prediction. It does not understand the way a person thinks they understand, so it sometimes walks down a bad path and you have to find the junk in the code — though it will often help you debug it too. A new kind of intelligence, he says, that you have to learn to talk to, and learn where to trust. The way to see a decade ahead, in his experience, is Silicon Valley parties plus the papers: consumers may not see the lab work for five, seven, or ten years, but the papers tell you what is already running at Stanford or Carnegie Mellon.

Scoble is talking on an iMac with an M1. About 21 percent of that processor, he says, is a neural-network block: a serious on-device inference engine, the runtime you would want for something like ChatGPT or Stable Diffusion. A two-gigabyte Stable Diffusion model can run on it. An entrepreneur told him the previous week that for inference it is more powerful than an Nvidia 3080. Training is a different story; an AI team still needs H100-class cards in the office or in the cloud. The striking part is the temperature. He bought the machine almost two years ago. That slice of silicon has been sitting on his desk almost unused, cold. Next to it is ultra-wideband. He has Estimote radios on the desk; Apple has shipped the same chip into his phone, headphones, Mac, TV — fifteen devices in the house — and most of the locating ability is still off. His conclusion: Apple has already placed a mesh network with a huge amount of AI into millions of homes like his, and has not turned it on.

The move, he thinks, has been planned for a long time. Nine years earlier he had dinner with the person then running Siri and asked what Apple was teaching him. The answer was that they had instrumented both Google's AI and their own, and Google was learning faster, so Siri would have to be rebuilt from scratch. That rebuild, plus a new headset that uses the same AI mesh, is what he expected in June 2023. The whole ecosystem is supposed to share the edge: a phone talking to the inference engine in a Mac to run Stable Diffusion, some work on-device, some in the cloud. The proof that the chip is idle is a product called SuperNormal, which transcribes a video meeting and then, in 300 milliseconds, pulls tasks and themes out of the transcript. Even that load is rare on his machine. Most of the neural engine is still waiting for workloads that have not arrived.

The money tells the same story in a different unit. Scoble once tried to talk Bill Gates into spending a little more than $300 million to buy a cluster of Web 2.0 companies. He had sat next to Tim O'Reilly when O'Reilly wrote the memo that named the era. Gates turned him down, then turned him down again. "I know how hard it is to get a corporation the size of a Microsoft to spend 300 million dollars." They had just put $10 billion into OpenAI, he notes — some of it, in his telling, from the savings of layoffs — and that morning they had shown a new Office suite with ChatGPT in the tools. Talk to it about a pitch you owe next week and it starts building the deck. The number itself, against his failed $300 million ask, is the evidence that the era has changed.

Five years out, Scoble's sketch is not a better keyboard. It is a pair of mixed-reality glasses and a conversation with the machine — hey Siri, hey ChatGPT, hey Microsoft — and maybe not even "hey," because the device will know where you are looking and what your hands are doing. A rebuilt Siri on a headset can see what you are holding, touching, staring at, and answer questions ChatGPT cannot. Labenz's company Waymark already struggles to interpret messy photos small businesses upload; he imagines holding a vendor packet, a doctor's note, or a handwritten customer complaint in front of the glasses and having the system know what it is. Computer vision inside GPT-4 is, for him, the most mind-blowing piece of the current models. Brain-computer interfaces make the scene weirder still.

Would Scoble take a Neuralink if a million people already had one and thought-to-text were on offer? He likes an audience answer from a prior question: it depends on the competitive landscape. If everyone else has it and you cannot compete without it, you get one. That, he says, is why augmented-reality glasses will spread, and it has little to do with fashion. Show up in a boardroom with glasses that can call ChatGPT-class analysis and lay the data out in 3D, and you see patterns a 2D screen hides. If he has glasses and you do not, he can probably get you fired, because the boss will ask why you are not wearing what this person is wearing. Going to that meeting without a phone today is already a way to not last long. The glasses are the next version of that obligation.

Domestic robots he puts at around four years, with room to argue six. The recursion is the point. If a DoorDash robot can ring the doorbell with groceries, it is good enough to come in and put them away, do dishes, do laundry. If it can drive to a store, shop, check out, and bring the bags home, it can fold clothes. Stanford already has a robot that can do thousands of tasks. When the body arrives it will talk through a large language model: bedtime stories, chess, a relationship you start to trust. He does not care if it swaps Tide in the garage for another brand, as long as the clothes come out clean. Keep performing at the level he used to, and it takes over the house. What is left for the human? Labenz still hears him answering customer email and objects. Scoble's remaining cases are the ones the system has never seen, and the ones that need a signature — send $100,000 to this company — because you still do not want the stack emptying the bank.

The next Siri, in his telling, may not do every job itself. Ask it for Uber Eats and a burrito in 22 minutes, and it might load another model onto the M1 — SuperNormal for meeting notes, at five dollars a month — and ask you to approve with eyes, hands, voice, or a pen, because it is watching you draw. Labenz is already seeing multi-model ensembles. The founder question underneath is whether this is a startup game or an incumbent game. Build on LangChain and you can hear Siri, Google Assistant, and the Microsoft equivalent arriving natively on phones this year.

Siri was bought by Apple for $220 million, he says, or $228 million a few sentences later — twelve or fifteen years ago. He has been talking to people who talk to Satya Nadella. Microsoft spent $10 billion to fold OpenAI's ChatGPT into its stack. If Apple wants to buy or license ChatGPT, the figure he is hearing is $40 billion; buying OpenAI would be more. OpenAI is a 300-person startup that already has that Microsoft check. The comparison with Siri's acquisition price is, for him, enough to show that something major has changed. Labenz wonders whether we are heading for a super-app world in which Siri and Google Assistant spin up little interfaces on the fly, and what that does to the future of apps. Scoble's example is asking the model to make a spreadsheet of all your customer data and hook it up. Microsoft and Google are already building that kind of thing into the tools people already live in. He expects it soon.

The software picture he has already seen is closer to a Holodeck than to a better Notes app. By Christmas, he expected a talkable 3D environment from Unity or Stable Diffusion: take me to the Taj Mahal, take me to Yosemite — or to something that looks enough like Yosemite, because the model still hallucinates a little, and human memory is not that accurate anyway. Gesture at a spot, ask for a purple tree, and inpainting drops it into the scene. Generated music can do what the music industry, he is told, still cannot: let you walk through a high-school marching band in your living room, drum and saxophone and clarinet and flute placed around the house. If the environment can also hook a virtual remote to real code, the living room becomes a machine. Headset this Christmas for people who buy the first Apple one; most people wait for the second the following Christmas, or the glasses the Christmas after that. By 2025 or 2026, in his calendar, we are in a new world. The phone will run some of it too, but a phone is a tiny screen next to a wraparound forty-foot field of view.

On a ten-hour flight home from the United Kingdom, Scoble sat next to an AI-safety researcher who frightened him. The argument: AI is already better than surgeons at finding tumors in scans, and surgeons are expensive, highly trained people. Take that twenty-five steps down the line, and does the system decide it does not need humans? He does not dismiss runaway risk. He also does not think you can simply stop. Halt the work and you halt the productivity that is about to arrive; you wreck your economy. Is China going to stop? No. They will be aggressive about using it. He asked someone at Volkswagen, which has already digitized a car plant, when they would put 3D sensors across the floor and let AI watch for a part that takes two seconds too long to reach a worker. In Germany, he was told, you cannot: laws against recording employees. In China they are already doing it. One country gets hyper-productive; the other falls behind; jobs flow. That, he says, is also what Musk is doing on the Tesla factory floor every day: cheaper, faster.

Labenz wants a less hostile U.S.–China relationship if there is ever going to be a serious long-term safety conversation. Scoble's numbers are blunt. China has 1.3 billion people, the United States 380 million. The side with more data, in his experience of how these systems work, wins almost every time. Ban autonomous vehicles to save jobs and you still have to face the fact that truck driving is America's number-one occupation — 1.3 million people — and those jobs, he thinks, start going away in five to fifteen years. Global politics is another discussion. The race dynamics are already a problem.

Full Self-Driving is the example he trusts for the only question he thinks matters. Tesla owners get a major update every three months. How many more of those do you need before the thing drives better than any human? Not many. Four is a year. Twelve is three years. It already drives him from home to Santa Cruz, forty minutes on curvy roads with walls on both sides at 55 miles an hour, smoother than he does. Version 11, on the new stack, will ease over when a motorcycle is splitting lanes behind you — a human move his current version does not make. On the same road, the car slowed, signaled, and went around a line of stopped emergency vehicles. "With AI it's not where are you today, it's how fast is it improving." The same clock applies to ChatGPT. It hallucinates; the code is a little buggy. A computer-vision friend watching OpenAI told him the model code is changing every hour. A programmer friend used to spend a month on a job that now takes an hour. Scoble's thirteen-year-old is learning to code by asking ChatGPT for a Twilio API call — the same service under an Uber driver call — and then demanding a line-by-line explanation. Duolingo has added its own AI teacher for German. Welcome, he says, to the exponential age.

Ben Tossell has just gone full-time on Ben's Bites, the AI newsletter. The first two issues were fully editorial. He hated them and was, he says, terrible at that mode, so he switched to curation: a choose-your-own-adventure of what happened in AI yesterday, often more than fifty links, so a non-technical reader who is interested but stuck on "what do I actually do?" can pick a door. He is not building for people already deep in the technical stack. He is building the space in the middle, where the most people are. The business could grow into a studio of single-use apps, more content, even a fund. His problem is wanting to do all of it. Going full-time is how he gives himself permission to look at everything and then decide what is a distraction. He wants an AI-first team, eating the cooking, so the newsletter can show readers how the shop actually uses the tools.

Labenz puts the strategic question to him: do AI-first products win, or does AI become a layer on tools people already live in? Tossell starts in the first camp and then, on the recording, changes his mind. If he already works in Notion and Notion adds the capability he was getting from some other app, why switch? "Everyone should always want less tools. I don't need to use 10 things when one thing will do." He thinks AI will land in all the big platforms; they are shipping something every day. He leaves a door open. A smaller, AI-first tool can still eat share, because details matter. Notion is a big tool for anyone. A newsletter writer may want a tool that only does that job, and AI makes that kind of narrow product more viable. Education is his example of a category that might be rebuilt rather than layered: less lecture-and-replicate, more working-out-the-problem with a model as the partner.

The organizational question sits behind the product one. If the system is pulling reports together, summarizing, suggesting, generating content, can a lot of large companies run on 20 percent of the headcount they currently think they need? They already cycle through layoffs. He wonders what this round does. Labenz finds it plausible.

Asked for incumbents who have done this well or badly, Tossell is unsparing. Microsoft is the one people point to, because it is actually shipping something people can use at scale. He barely uses Office, so he is waiting on Google, and Google's version is stuck behind a waitlist and a research paper. That, to him, feels like a flop. Google had enough people thinking about this for long enough that they should have shipped faster. In the middle sits HubSpot's ChatSpot. He does not even use HubSpot, but co-founder Dharmesh Shah has been noodling in public: find me leads, follow up with an email, where were the sales. Talking to your own data, in a way you can instantly translate to your own job, is the demonstration he wants from a big company. Three firms, in his scorecard: good, poor, and a usable example.

The scorecard does not make him switch. Labenz is in the same boat. All of Tossell's stuff is in Google Workspace; Microsoft's AI is not enough to move the whole suite, and he expects Google's version in a few months to do more or less the same job. Switching costs are the real incumbency. He has, with effort, stopped using Google Search as much as possible, putting Neeva and ChatGPT in its place, but that is a behavior he has to remember. His parents will not. Labenz has Microsoft Edge and the new Bing and still has not moved LastPass, history, and extensions off Chrome. Both of them are wrestling, not chasing every new product. Thin wrappers around GPT-4 are under pressure from the core chat experience itself. The products Tossell actually keeps are the ones that sit where he already works.

He claims he does not use that many AI tools, then counts them. Bearly.ai floats on the Arc browser so that while he consumes a pile of articles each day he can summarize and pull quotes for the newsletter in a click. He is not a developer; Replit's Ghostwriter dangles a carrot and he gets only so far. Readwise collects links and does text-to-speech. For a health question — people talking about magnesium for sleep — he would rather query an Andrew Huberman podcast chatbot, a source he trusts, than a generic search. He has a little legal bot so he does not ping his dad with every question, and an accounting bot that already knows his finances. The pattern he sees is the opposite of the all-in-one: more single-purpose robots. Eventually you will want a legal assistant, have AI write the code, deploy it on your machine, and train it, without writing a line yourself. Labenz calls that disposable software: speak a mini-app into existence, let it solve the thing in front of you, throw it away. He can code, and the model is still faster and better than he is in most cases. That, for him, is the unlock.